Your rig

RTX 4000 Ada · 20 GB

16 GB system RAM · Ada Lovelace · 360 GB/s · 18.4 GB usable VRAM
Edit hardware Export as JSON

Context length

KV cache at f16, 8K tokens

Quality floor
Q8_0 — near-lossless Q6_K Q5_K_M Q4_K_M — recommended Q3_K_M Q2_K — damaged
KV cache type
System RAM (GB)
Runtime
Ollamallama.cppLM StudiovLLM

4 of 5 runtimes drive this device.

Allow CPU offload Include partial-offload fits

Offloaded layers run at 60 GB/s host bandwidth — expect single-digit tokens per second.

Which one should I actually run?

Best models for your RTX 4000 Ada

Updated 8 Oct 2026 1 new this month

What do you want to do? Each row is the best-ranked model on that use case's shortlist that fits at Q3_K_M and 8K. The ordering is editorial; the sizes and speeds are computed. How the shortlists work.

RecommendationModelWhyQuant · totalTok/s est.
Best overall Qwen3.8 27B
Alibaba · Apache 2.0
Best capability per gigabyte: a current-generation dense 27B with vision and 262K context. Q3_K_M · 13.8 GB 17 Runs great
Best for coding Qwen3.8 27B
Alibaba · Apache 2.0
Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. Q3_K_M · 13.8 GB 17 Runs great
Best reasoning Qwen3.8 27B
Alibaba · Apache 2.0
Thinking mode plus the best agentic reasoning at 24 GB. Q3_K_M · 13.8 GB 17 Runs great
Best for writing Gemma 4 26B-A4B
Google · Apache 2.0
Gemma 4 prose at MoE speed, with room for a long draft in context. Q3_K_M · 13.2 GB 48 Runs great
Best vision Qwen3.8 27B
Alibaba · Apache 2.0
Image and video in, OSWorld 84 — the strongest local vision-language model at 24 GB. Q3_K_M · 13.8 GB 17 Runs great
Best for agents Qwen3.8 27B
Alibaba · Apache 2.0
Computer use and tool calling are what 3.8 was trained for (OSWorld 84). Q3_K_M · 13.8 GB 17 Runs great
Best translation Qwen3.8 27B
Alibaba · Apache 2.0
119 languages, and Qwen translations read as idiomatic rather than literal. Q3_K_M · 13.8 GB 17 Runs great
Fastest good model Gemma 4 26B-A4B
Google · Apache 2.0
3.8B active and a small file — the quickest Gemma 4 with vision. Q3_K_M · 13.2 GB 48 Runs great
Best long context Nemotron 3.5 Lightning 30B-A3B
NVIDIA · OpenMDW-1.1
262K, and a Mamba hybrid whose cache barely grows. Q3_K_M · 15.7 GB
at 128K context
58 Tight fit

Everything that fits

Runs great

49 of 80 · fits with headroom for a longer prompt
ModelParamsQuantWeights +KV 8KTotal VRAM headroom Tok/sRuntime
Nemotron 3.5 Lightning 30B-A3B 31.6B MoE / 3.2B act Q3_K_M 14.4 GB 0.05 GB 15.0 GB
3.4 GB free
58 Ollama Run
GLM-4.7-Flash 30B-A3B 31.2B MoE / 3B act Q3_K_M 14.2 GB 0.41 GB 15.2 GB
3.2 GB free
61 Ollama Run
Qwen3 30B-A3B 30.5B MoE / 3.3B act Q3_K_M 13.9 GB 0.75 GB 15.2 GB
3.2 GB free
56 Ollama Run
Qwen3 Coder 30B-A3B 30.5B MoE / 3.3B act Q3_K_M 13.9 GB 0.75 GB 15.2 GB
3.2 GB free
56 Ollama Run
Muse Glimmer 30B 29.8B Q3_K_M 13.6 GB 0.18 GB 14.3 GB
4.1 GB free
16 Ollama Run
Qwen3.8 27B 27.8B Q3_K_M 12.7 GB 0.50 GB 13.8 GB
4.6 GB free
17 Ollama Run
Qwen3.6 27B 27.8B Q3_K_M 12.7 GB 0.50 GB 13.8 GB
4.6 GB free
17 Ollama Run
Gemma 3 27B 27.4B Q3_K_M 12.5 GB 1.03 GB 14.1 GB
4.3 GB free
17 Ollama Run
Gemma 4 26B-A4B 26.5B MoE / 3.8B act Q3_K_M 12.1 GB 0.51 GB 13.2 GB
5.2 GB free
48 Ollama Run
Devstral Small 2 24B 24B Q3_K_M 10.9 GB 1.25 GB 12.8 GB
5.6 GB free
20 llama.cpp Run
Mistral Small 3.2 24B 23.6B Q3_K_M 10.7 GB 1.25 GB 12.6 GB
5.8 GB free
20 Ollama Run
gpt-oss 20B 20.9B MoE / 3.6B act MXFP4 10.8 GB 0.19 GB 11.6 GB
6.8 GB free
45 Ollama Run
Qwen3 14B 14.8B Q3_K_M 6.7 GB 1.25 GB 8.6 GB
9.8 GB free
32 Ollama Run
DeepSeek-R1-Distill-Qwen 14B 14.8B Q3_K_M 6.7 GB 1.50 GB 8.8 GB
9.6 GB free
32 Ollama Run
Qwen2.5-Coder 14B 14.8B Q3_K_M 6.7 GB 1.50 GB 8.8 GB
9.6 GB free
32 Ollama Run
Phi-4 14B 14.7B Q3_K_M 6.7 GB 1.56 GB 8.9 GB
9.5 GB free
33 Ollama Run
Ministral 3 14B 13.9B Q3_K_M 6.3 GB 1.25 GB 8.2 GB
10.2 GB free
34 Ollama Run
Gemma 3 12B 12.2B Q3_K_M 5.6 GB 0.81 GB 7.0 GB
11.4 GB free
39 Ollama Run
Mistral NeMo 12B 12.2B Q3_K_M 5.6 GB 1.25 GB 7.4 GB
11.0 GB free
39 Ollama Run
Gemma 4 12B 12B Q3_K_M 5.5 GB 0.81 GB 6.9 GB
11.5 GB free
40 Ollama Run
Qwen3.5 9B 9.65B Q3_K_M 4.4 GB 0.25 GB 5.2 GB
13.2 GB free
50 Ollama Run
Ornith 1.5 9B 9.41B Q3_K_M 4.3 GB 0.25 GB 5.1 GB
13.3 GB free
51 Ollama Run
Ministral 3 8B 8.92B Q3_K_M 4.1 GB 1.06 GB 5.7 GB
12.7 GB free
54 Ollama Run
Granite 4.1 8B 8.79B Q3_K_M 4.0 GB 1.25 GB 5.9 GB
12.5 GB free
54 Ollama Run
LFM2.5 8B-A1B 8.47B MoE / 1.5B act Q3_K_M 3.9 GB 0.09 GB 4.5 GB
13.9 GB free
123 Ollama Run
Fara 7B 8.29B Q3_K_M 3.8 GB 0.44 GB 4.8 GB
13.6 GB free
58 llama.cpp Run
Qwen3 8B 8.19B Q3_K_M 3.7 GB 1.13 GB 5.5 GB
12.9 GB free
58 Ollama Run
Llama 3.1 8B Instruct 8.03B Q3_K_M 3.7 GB 1.00 GB 5.3 GB
13.1 GB free
60 Ollama Run
Gemma 4 E4B 8.0B Q3_K_M 3.6 GB 0.14 GB 4.4 GB
14.0 GB free
60 Ollama Run
Ling 3.0 Tiny 7.9B-A1.3B 7.9B MoE / 1.3B act Q3_K_M 3.6 GB 0.05 GB 4.2 GB
14.2 GB free
142 llama.cpp Run
DeepSeek-R1-Distill-Qwen 7B 7.62B Q3_K_M 3.5 GB 0.44 GB 4.5 GB
13.9 GB free
63 Ollama Run
Qwen2.5-Coder 7B 7.62B Q3_K_M 3.5 GB 0.44 GB 4.5 GB
13.9 GB free
63 Ollama Run
Olmo 3 7B Instruct 7.3B Q3_K_M 3.3 GB 2.50 GB 6.4 GB
12.0 GB free
66 Ollama Run
Mistral 7B Instruct v0.3 7.25B Q3_K_M 3.3 GB 1.00 GB 4.9 GB
13.5 GB free
66 Ollama Run
Gemma 4 E2B 5.1B Q3_K_M 2.3 GB 0.07 GB 3.0 GB
15.4 GB free
94 Ollama Run
Qwen3.5 4B 4.66B Q3_K_M 2.1 GB 0.25 GB 3.0 GB
15.4 GB free
103 Ollama Run
Gemma 3 4B 4.3B Q3_K_M 2.0 GB 0.30 GB 2.9 GB
15.5 GB free
111 Ollama Run
Qwen3 4B 4.02B Q3_K_M 1.8 GB 1.13 GB 3.6 GB
14.8 GB free
119 Ollama Run
Ministral 3 3B 3.85B Q3_K_M 1.8 GB 0.81 GB 3.2 GB
15.2 GB free
124 Ollama Run
Phi-4-mini 3.8B 3.84B Q3_K_M 1.7 GB 1.00 GB 3.3 GB
15.1 GB free
125 Ollama Run
Granite 4.1 3B 3.4B Q3_K_M 1.5 GB 0.63 GB 2.8 GB
15.6 GB free
141 Ollama Run
Llama 3.2 3B Instruct 3.21B Q3_K_M 1.5 GB 0.88 GB 2.9 GB
15.5 GB free
149 Ollama Run
LFM2.5 2.6B 2.7B Q3_K_M 1.2 GB 0.13 GB 2.0 GB
16.4 GB free
177 llama.cpp Run
Qwen3.5 2B 2.27B Q3_K_M 1.0 GB 0.09 GB 1.7 GB
16.7 GB free
211 Ollama Run
Qwen3 1.7B 1.72B Q3_K_M 0.8 GB 0.88 GB 2.3 GB
16.1 GB free
278 Ollama Run
Llama 3.2 1B Instruct 1.24B Q3_K_M 0.6 GB 0.25 GB 1.4 GB
17.0 GB free
386 Ollama Run
Gemma 3 1B 1.0B Q3_K_M 0.5 GB 0.04 GB 1.1 GB
17.3 GB free
479 Ollama Run
Qwen3.5 0.8B 0.87B Q3_K_M 0.4 GB 0.09 GB 1.1 GB
17.3 GB free
550 Ollama Run
Qwen3 0.6B 0.6B Q3_K_M 0.3 GB 0.88 GB 1.7 GB
16.7 GB free
798 Ollama Run

Runs — tight

9 models · fits at 8K, drop context or quant to go longer
ModelParamsQuantWeights +KV 8KTotal VRAM headroom Tok/sRuntime
Qwen3.6 35B-A3B 35.9B MoE / 3.3B act Q3_K_M 16.3 GB 0.16 GB 17.1 GB
1.3 GB free
56 llama.cpp Run
Ornith 1.5 35B-A3B 35.9B MoE / 3.3B act Q3_K_M 16.3 GB 0.16 GB 17.1 GB
1.3 GB free
56 llama.cpp Run
LLM-jp 4 33B Thinking 33.2B Q3_K_M 15.1 GB 2.00 GB 17.7 GB
0.7 GB free
14 llama.cpp Run
Qwen3 32B 32.8B Q3_K_M 14.9 GB 2.00 GB 17.5 GB
0.9 GB free
15 llama.cpp Run
DeepSeek-R1-Distill-Qwen 32B 32.8B Q3_K_M 14.9 GB 2.00 GB 17.5 GB
0.9 GB free
15 llama.cpp Run
Qwen2.5-Coder 32B 32.8B Q3_K_M 14.9 GB 2.00 GB 17.5 GB
0.9 GB free
15 llama.cpp Run
Olmo 3.1 32B Instruct 32.2B Q3_K_M 14.7 GB 1.25 GB 16.5 GB
1.9 GB free
15 llama.cpp Run
Gemma 4 31B 31.3B Q3_K_M 14.2 GB 2.03 GB 16.9 GB
1.5 GB free
15 llama.cpp Run
Granite 4.1 30B 28.9B Q3_K_M 13.2 GB 2.00 GB 15.8 GB
2.6 GB free
17 llama.cpp Run

Won't fit in VRAM

22 models · some are possible with CPU offload
ModelParamsQuantWeights +KV 8KTotal Over budget Tok/sRuntime
Kimi K3 2.8T-A104B 2780B MoE / 104B act Q3_K_M 1265.4 GB 0.21 GB 1266.2 GB +1247.8 GB ~0.3 llama.cpp Why
Qwen3.8 2.4T-A95B 2446B MoE / 95B act Q3_K_M 1113.4 GB 0.72 GB 1114.7 GB +1096.3 GB ~0.3 llama.cpp Why
DeepSeek V4 Pro 1.6T-A49B 1650B MoE / 49B act Q3_K_M 751.1 GB 0.54 GB 752.2 GB +733.8 GB ~0.6 llama.cpp Why
Kimi K2.6 1T-A32B 1027B MoE / 32B act Q3_K_M 467.5 GB 0.54 GB 468.6 GB +450.2 GB ~1.0 llama.cpp Why
GLM-5.2 744B-A40B 753B MoE / 40B act Q3_K_M 342.8 GB 0.69 GB 344.0 GB +325.6 GB ~0.8 llama.cpp Why
GLM-5.3 744B-A40B 753B MoE / 40B act Q3_K_M 342.8 GB 0.69 GB 344.0 GB +325.6 GB ~0.8 llama.cpp Why
DeepSeek-R1 671B 671B MoE / 37B act Q3_K_M 305.4 GB 0.54 GB 306.6 GB +288.2 GB ~0.9 llama.cpp Why
DeepSeek V4.1 Flash 552B-A16B New 552B MoE / 16B act Q3_K_M 251.3 GB 0.04 GB 251.9 GB +233.5 GB ~2.0 vLLM Why
Qwen3.5 397B-A17B 403B MoE / 17B act Q3_K_M 183.4 GB 0.23 GB 184.3 GB +165.9 GB ~2.0 llama.cpp Why
Ornith 1.5 397B-A17B 397B MoE / 17B act Q3_K_M 180.7 GB 0.23 GB 181.5 GB +163.1 GB ~2.0 llama.cpp Why
GLM-5.3-Flash 320B-A18B 321B MoE / 18B act Q3_K_M 146.1 GB 0.09 GB 146.8 GB +128.4 GB ~1.9 llama.cpp Why
DeepSeek V4 Flash 284B-A13B 284B MoE / 13B act Q3_K_M 129.3 GB 0.38 GB 130.3 GB +111.9 GB ~2.7 llama.cpp Why
Qwen3 235B-A22B 235B MoE / 22B act Q3_K_M 107.0 GB 1.47 GB 109.0 GB +90.6 GB ~1.6 llama.cpp Why
Qwen3.8-Flash-Next 180B-A6B 180B MoE / 6B act Q3_K_M 81.9 GB 0.19 GB 82.7 GB +64.3 GB ~6.2 llama.cpp Why
Qwen3.5 122B-A10B 125B MoE / 10B act Q3_K_M 56.9 GB 0.19 GB 57.7 GB +39.3 GB ~4.1 llama.cpp Why
Nemotron 3 Super 120B-A12B 124B MoE / 12B act Q3_K_M 56.4 GB 0.06 GB 57.1 GB +38.7 GB ~3.5 llama.cpp Why
Ling 3.0 Flash 124B-A5B 124B MoE / 5.1B act Q3_K_M 56.4 GB 0.06 GB 57.1 GB +38.7 GB ~8.2 llama.cpp Why
Mistral Small 4 119B-A6B 119B MoE / 6.5B act Q3_K_M 54.2 GB 0.18 GB 54.9 GB +36.5 GB ~6.5 llama.cpp Why
gpt-oss 120B 117B MoE / 5.1B act MXFP4 60.5 GB 0.29 GB 61.4 GB +43.0 GB ~7.0 llama.cpp Why
Llama 4 Scout 109B-A17B 109B MoE / 17B act Q3_K_M 49.6 GB 1.50 GB 51.7 GB +33.3 GB ~2.5 llama.cpp Why
Llama 3.3 70B Instruct 70.6B Q3_K_M 32.1 GB 2.50 GB 35.2 GB +16.8 GB ~1.9 llama.cpp Why
DeepSeek-R1-Distill-Llama 70B 70.6B Q3_K_M 32.1 GB 2.50 GB 35.2 GB +16.8 GB ~1.9 llama.cpp Why

Totals = quantised weights + f16 KV cache at 8K tokens + 0.6 GB runtime overhead, against 18.4 GB usable VRAM (20 GB less display and driver reserve). Tokens per second are estimated from 360 GB/s peak memory bandwidth at batch 1 — not measured. Read the method.


Same memory, different speed

These devices have the same 18.4 GB usable, so they run exactly the models above. Generation speed scales with memory bandwidth.

DeviceBandwidthSpeed vs yours
Radeon RX 7900 XT
RDNA 3 · 2022
800 GB/s 2.22×

What more memory unlocks

At Q4_K_M and 8K context this device fits 45 of 80 models. Each step up, and the largest models it adds:

Every device ranked by what it can run →